Pith. sign in

REVIEW 2 major objections 2 minor 28 references

Active Source-free Domain Adaptation in Open-set Medical Image Segmentation via Decomposed Uncertainty and Prototype Discrepancy

T0 review · 2 major / 2 minor · reviewed 2026-06-27 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read An active source-free open-set adaptation method uses decomposed uncertainty and prototype discrepancy to select samples for medical image segmentation without source data.

desk verdict The paper introduces a coherent active selection strategy for source-free open-set medical segmentation adaptation but the abstract supplies no numbers or ablations to check whether the components deliver the claimed gains. read the letter →

arxiv 2606.08749 v1 pith:DIP2OB2B submitted 2026-06-07 eess.IV

Jin Yang , Yichi Zhang , Peijie Qiu , Xiaobing Yu This is my paper
classification eess.IV
keywords activedomainadaptationsource-freeopen-setsegmentationmedicalimageuncertaintyestimationprototypediscrepancyself-training
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper aims to adapt deep learning segmentation models to new medical imaging domains that contain unseen classes, while obeying privacy rules that block access to the original source data. It develops an active learning strategy that queries a small set of target samples based on class-aware decomposed uncertainty, which captures both data noise and model uncertainty through test-time augmentation, plus class-agnostic prototype discrepancy that tracks domain shifts and sample diversity. A target-refined self-training step then creates pseudo labels for the remaining samples so they can be used in semi-supervised training. The authors claim this combination yields higher segmentation accuracy than prior adaptation techniques on cross-domain volumetric tasks. If the approach holds, models could be updated in clinical settings that introduce new pathologies without violating data-sharing restrictions.

What carries the argument

The active open-set query strategy that combines Class-aware Decomposed Uncertainty (CDU) measured by test-time augmentation and Class-agnostic Prototype Discrepancy (CPD) to choose samples, paired with Target-refined Self-training for pseudo-label generation.

What would settle it

An experiment on the same cross-domain open-set volumetric tasks where random sample selection or existing adaptation methods achieve equal or higher Dice scores than the proposed selection and self-training pipeline.

Watch

Extended reading notes

Core claim

The ASFOSDA method is the first to apply active learning to source-free open-set domain adaptation for medical image segmentation by selecting informative target samples via Class-aware Decomposed Uncertainty and Class-agnostic Prototype Discrepancy and then applying Target-refined Self-training to generate pseudo labels for unselected samples.

Load-bearing premise

The uncertainty and discrepancy calculations will reliably pick the samples that most improve adaptation performance, and the self-training will produce pseudo labels accurate enough to boost results on unselected data.

Editorial extensions

If this is right

  • Adaptation becomes feasible when source data cannot be shared due to privacy regulations.
  • Models can segment target domains that include anatomical structures or pathologies absent from the source set.
  • Active selection lowers the number of target annotations needed while still improving over passive or closed-set methods.
  • Combining the selected labeled samples with pseudo-labeled ones creates a stronger semi-supervised training signal.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same selection criteria could be tested in non-medical domains that also face privacy barriers and open-set conditions.
  • Replacing the current uncertainty estimator with other test-time methods might further reduce the number of queries required.
  • Applying the framework to 2D slice data or different imaging modalities would check whether the gains generalize beyond volumetric scans.
  • The approach suggests a route for blending active learning with self-supervised signals in other medical AI tasks.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes ASFOSDA, the first active source-free open-set domain adaptation method for medical image segmentation. It selects the most informative target samples via Class-aware Decomposed Uncertainty (CDU, using test-time augmentation to capture aleatoric and epistemic uncertainty) and Class-agnostic Prototype Discrepancy (CPD, measuring cross- and self-domain discrepancy). A Target-refined Self-training strategy then generates pseudo-labels for unselected samples, enabling semi-supervised training on the combined set. The method is evaluated on cross-domain open-set volumetric medical image segmentation tasks and reported to outperform state-of-the-art adaptation approaches.

Significance. If the reported gains hold under rigorous validation, the work is significant for addressing the practical intersection of source-free constraints (privacy), open-set conditions (novel anatomical/pathological classes), and active learning (limited labeling budget) in clinical medical imaging. The pipeline is coherent, and the explicit use of decomposed uncertainty plus prototype discrepancy for selection, together with the self-training refinement, provides a concrete, falsifiable approach that could be extended to other volumetric tasks.

major comments (2)
  1. [§4] §4 (Experiments): The central claim of outperformance over SOTA methods is load-bearing, yet the provided description supplies no quantitative metrics (Dice, HD95, etc.), ablation tables isolating CDU vs. CPD contributions, dataset specifications (modalities, number of volumes, open-set class splits), or statistical significance tests; without these, the data-to-claim link cannot be verified.
  2. [Method] Method section, Target-refined Self-training paragraph: The strategy assumes that pseudo-labels generated for unselected samples remain sufficiently accurate in the presence of open-set classes; this assumption is load-bearing for the semi-supervised stage, but no analysis of pseudo-label error rates or failure modes on novel classes is referenced.
minor comments (2)
  1. [Abstract] Abstract: The sentence claiming 'it outperformed state-of-the-art adaptation methods' should be accompanied by at least one concrete metric or reference to the results table for immediate clarity.
  2. [Method] Notation: CDU and CPD are introduced with full names but the subsequent text occasionally uses the acronyms without re-definition; a short notation table or consistent first-use expansion would improve readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's significance. We address each major comment below.

read point-by-point responses

  1. Referee: [§4] §4 (Experiments): The central claim of outperformance over SOTA methods is load-bearing, yet the provided description supplies no quantitative metrics (Dice, HD95, etc.), ablation tables isolating CDU vs. CPD contributions, dataset specifications (modalities, number of volumes, open-set class splits), or statistical significance tests; without these, the data-to-claim link cannot be verified.

    Authors: The full manuscript contains tables reporting Dice and HD95 metrics across the evaluated volumetric tasks along with comparisons to prior methods. However, we agree that dedicated ablation tables isolating CDU versus CPD, fuller dataset specifications, and statistical significance tests are not sufficiently highlighted. We will add these elements in the revised version to make the empirical support fully explicit. revision: partial

  2. Referee: [Method] Method section, Target-refined Self-training paragraph: The strategy assumes that pseudo-labels generated for unselected samples remain sufficiently accurate in the presence of open-set classes; this assumption is load-bearing for the semi-supervised stage, but no analysis of pseudo-label error rates or failure modes on novel classes is referenced.

    Authors: We acknowledge that the current manuscript does not provide explicit quantitative analysis of pseudo-label error rates specifically on novel (open-set) classes. While the overall pipeline is designed to mitigate this via the active selection step, we agree an additional analysis would strengthen the claims. We will incorporate such an analysis (e.g., error-rate tables or failure-case discussion) in the revision. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified

full rationale

The provided abstract and method description outline an empirical pipeline for ASFOSDA using CDU (via TTA for uncertainty) and CPD (for discrepancy) to select samples, plus target-refined self-training for pseudo-labels. No equations, derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear. Claims of outperformance rest on experimental evaluation rather than any reduction of outputs to inputs by construction. This is the common case of a self-contained empirical contribution with no circular steps.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; CDU and CPD are introduced as new selection mechanisms but their internal formulations and any tunable thresholds are not stated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Source-free Domain Adaptation in Open-set Medical Image Segmentation via Decomposed Uncertainty and Prototype Discrepancy." pith.science (2026). https://pith.science/paper/DIP2OB2B

@misc{pith2026260608749,
  author       = {Pith},
  title        = {Pith review of: Active Source-free Domain Adaptation in Open-set Medical Image Segmentation via Decomposed Uncertainty and Prototype Discrepancy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DIP2OB2B}},
  note         = {Machine review of arXiv:2606.08749}
}
read the original abstract

Deep learning (DL) methods are challenged to demonstrate robust performance across different segmentation datasets due to domain shifts, but active domain adaptation techniques enhance their generalization performance by querying a few samples from target domains for adaptation training. However in clinical practice, target domains often include private classes of new anatomical structures or pathologies that are not presented in the source data, and existing methods implement closed-set segmentation where source and target domains have the same segmentation classes. Additionally, source data are often inaccessible during adaptation due to strict data privacy regulations. To address these limitations, we propose an Active Source-free Open-set Domain Adaptation (ASFOSDA) method which is the first work to implement active learning for adapting DL models in open-set medical image segmentation without the access to source data. This method employs an active open-set query strategy to select the most informative target samples for training models based on Class-aware Decomposed Uncertainty (CDU) and Class-agnostic Prototype Discrepancy (CPD). CDU measures sample aleatoric uncertainty and model epistemic uncertainty by employing test time augmentation in stochastic processes. CPD measures cross-domain and self-domain discrepancy for selecting diverse samples. Subsequently, to boost the adaptation performance by enhancing training samples, a Target-refined Self-training strategy is proposed to generate high-quality pseudo labels for unselected samples, thus combining them with labeled samples for a semi-supervised training. We evaluated our method on cross-domain open-set volumetric medical image segmentation tasks, and it outperformed state-of-the-art adaptation methods.

Figures

Figures reproduced from arXiv: 2606.08749 by the authors.

Figure 1
Figure 1. Active Source-free Open-set Domain Adaptation in medical image [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison among results of SFDA, ASFDA, and our methods by adapting U-Net to (A) AMOS CT and (B) AMOS MR datasets. and cross entropy loss was used in supervised training and self-training. Networks were initialized on source data for 1000 epochs and trained on target data for 500 epochs. The AdamW was utilized for optimization. The initial learning rate was set to 0.001 and decayed in a polynomial sched… view at source ↗
Figure 3
Figure 3. Ablation studies of key hyper-parameters and components. (A) Impact of differ￾ent target budgets on U-Net adaptation. (B) Impact of different number of prototypes and candidates K using 5% budget for U-Net FLARE-to-AMOS adaptation. (C) Per￾formance comparison of uncertainty methods in U-Net MSD-to-FLARE adaptation [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

28 extracted references · 1 canonical work pages

  1. [1]

    Nature communications13(1), 4128 (2022)

    Antonelli, M., Reinke, A., Bakas, S., Farahani, K., Kopp-Schneider, A., Landman, B.A., Litjens, G., Menze, B., Ronneberger, O., Summers, R.M., et al.: The medical segmentation decathlon. Nature communications13(1), 4128 (2022)

    2022

  2. [2]

    Medical Image Analysis82, 102617 (2022)

    Bateson, M., Kervadec, H., Dolz, J., Lombaert, H., Ayed, I.B.: Source-free domain adaptation for image segmentation. Medical Image Analysis82, 102617 (2022)

    2022

  3. [3]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Chen,C.,Liu,Q.,Jin,Y.,Dou,Q.,Heng,P.A.:Source-freedomainadaptivefundus image segmentation with denoised pseudo-labeling. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 225–235. Springer (2021)

    2021

  4. [4]

    Medical Image Analysis90, 102958 (2023)

    Gaillochet, M., Desrosiers, C., Lombaert, H.: Active learning for medical image segmentation with stochastic batches. Medical Image Analysis90, 102958 (2023)

    2023

  5. [5]

    IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

    Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

    2021

  6. [6]

    In: International MICCAI brainlesion workshop

    Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

    2021

  7. [7]

    Advances in neural information processing systems 35, 36722–36732 (2022)

    Ji, Y., Bai, H., Ge, C., Yang, J., Zhu, Y., Zhang, R., Li, Z., Zhanng, L., Ma, W., Wan, X., et al.: Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation. Advances in neural information processing systems 35, 36722–36732 (2022)

    2022

  8. [8]

    IEEE transactions on pattern analysis and machine intelligence28(8), 1251–1261 (2006)

    Li, M., Sethi, I.K.: Confidence-based active learning. IEEE transactions on pattern analysis and machine intelligence28(8), 1251–1261 (2006)

    2006

Show all 28 references
  1. [9]

    In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI)

    Liu, X., Araki, K., Harada, S., Yoshizawa, A., Terada, K., Kurata, M., Nakajima, N., Abe, H., Ushiku, T., Bise, R.: Cluster entropy: Active domain adaptation in pathological image segmentation. In: 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI). pp. 1–5. I...

    2023

  2. [10]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Luo, Z., Luo, X., Gao, Z., Wang, G.: An uncertainty-guided tiered self-training framework for active source-free domain adaptation in prostate segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 107–117. Springer (2024)

    2024

  3. [11]

    Medical Image Analysis82, 102616 (2022)

    Ma, J., Zhang, Y., Gu, S., An, X., Wang, Z., Ge, C., Wang, C., Zhang, F., Wang, Y., Xu, Y., et al.: Fast and low-gpu-memory abdomen ct organ segmentation: the flare challenge. Medical Image Analysis82, 102616 (2022)

    2022

  4. [12]

    In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

    Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

    2015

  5. [13]

    In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1aIuk-RW

    Sener, O., Savarese, S.: Active learning for convolutional neural networks: A core- set approach. In: International Conference on Learning Representations (2018), https://openreview.net/forum?id=H1aIuk-RW

    2018

  6. [14]

    In: Medical Imaging with Deep Learning (2024),https://openreview.net/forum?id=Ym30gCtKqN

    VS, V., Valanarasu, J.M.J., Patel, V.M.: Target and task specific source-free domain adaptive image segmentation. In: Medical Imaging with Deep Learning (2024),https://openreview.net/forum?id=Ym30gCtKqN

    2024

  7. [15]

    In: 2014 International joint conference on neural networks (IJCNN)

    Wang, D., Shang, Y.: A new active labeling method for deep learning. In: 2014 International joint conference on neural networks (IJCNN). pp. 112–119. IEEE (2014)

    2014

  8. [16]

    IEEE Transactions on Medical Imaging (2024)

    Wang, H., Chen, J., Zhang, S., He, Y., Xu, J., Wu, M., He, J., Liao, W., Luo, X.: Dual-reference source-free active domain adaptation for nasopharyngeal carci- noma tumor segmentation across multiple hospitals. IEEE Transactions on Medical Imaging (2024)

    2024

  9. [17]

    Pattern Recognition114, 107870 (2021)

    Wang, W., Wang, H., Ran, Z.Y., He, R.: Learning robust feature transformation for domain adaptation. Pattern Recognition114, 107870 (2021)

    2021

  10. [18]

    Medical image analysis55, 88–102 (2019)

    Wang, Y., Zhou, Y., Shen, W., Park, S., Fishman, E.K., Yuille, A.L.: Abdomi- nal multi-organ segmentation with organ-attention networks and statistical fusion. Medical image analysis55, 88–102 (2019)

    2019

  11. [19]

    In: Proceedings of the AAAI conference on artificial intelligence

    Xie,B.,Yuan,L.,Li,S.,Liu,C.H.,Cheng,X.,Wang,G.:Activelearningfordomain adaptation: An energy-based approach. In: Proceedings of the AAAI conference on artificial intelligence. vol. 36, pp. 8708–8716 (2022)

    2022

  12. [20]

    Medical Image Analysis79, 102457 (2022)

    Yang, C., Guo, X., Chen, Z., Yuan, Y.: Source free domain adaptation for medical image segmentation with fourier style mining. Medical Image Analysis79, 102457 (2022)

    2022

  13. [21]

    arXiv preprint arXiv:2509.10784 (2025)

    Yang, J., Marcus, D.S., Sotiras, A.: Adapting medical vision foundation models for volumetric medical image segmentation via active learning and selective semi- supervised fine-tuning. arXiv preprint arXiv:2509.10784 (2025)

    2025 arXiv

  14. [22]

    Biomedical Signal Processing and Control113, 108837 (2026)

    Yang, J., Qiu, P., Zhang, Y., Marcus, D.S., Sotiras, A.: D-net: Dynamic large kernel with dynamic feature fusion for volumetric medical image segmentation. Biomedical Signal Processing and Control113, 108837 (2026)

    2026

  15. [23]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Yang, J., Yu, X., Qiu, P., Marcus, D., Sotiras, A.: Active source-free cross-domain and cross-modality adaptation for volumetric medical image segmentation by im- age sensitivity and organ heterogeneity sampling. In: International Conference on Medical Image Computing and Comp...

    2025

  16. [24]

    In: International conference on medical image computing and computer-assisted intervention

    Yang,L.,Zhang,Y.,Chen,J.,Zhang,S.,Chen,D.Z.:Suggestiveannotation:Adeep active learning framework for biomedical image segmentation. In: International conference on medical image computing and computer-assisted intervention. pp. 399–407. Springer (2017) ASFOSDA for Medical Imag...

    2017

  17. [25]

    Neural Net- works p

    Yang, Z., Li, Y., Liu, Y.: Continual source-free active domain adaptation for na- sopharyngeal carcinoma tumor segmentation across multiple hospitals. Neural Net- works p. 107869 (2025)

    2025

  18. [26]

    Medical Image Analysis102, 103543 (2025)

    Zhang, T., Li, K., Gu, S., Heng, P.A.: Enhancing source-free domain adaptation in medical image segmentation via regulated model self-training. Medical Image Analysis102, 103543 (2025)

    2025

  19. [27]

    Radiology Advances2(1), umae035 (2025)

    Zhou, A., Liu, Z., Tieu, A., Patel, N., Sun, S., Yang, A., Choi, P., Lee, H.C., Tordjman, M., Deyer, L., et al.: Mrannotator: multi-anatomy and many-sequence mri segmentation of 44 structures. Radiology Advances2(1), umae035 (2025)

    2025

  20. [28]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Zhou, T., Yang, J., Cui, L., Zhang, N., Chai, S.: Sbc-al: Structure and boundary consistency-based active learning for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 283–293. Springer (2024)

    2024

Pith tools

Reviewed June 27, 2026 · model on record in the stance chip above.